G2VD: Generalizable AI-Generated Video Detection via Counterfactual Intervention and Causal Disentanglement
G2VD integrates counterfactual intervention and causal disentanglement, achieving over 90% accuracy in cross-domain AI video forgery detection.
Key Findings
Methodology
G2VD employs a VAE-based counterfactual sample generation (CFIPipeline), aligning frequency and pixel domains to weaken spurious correlations. Based on a structural causal model (SCM), it introduces two domain-anchored branches—causal and non-causal—using HSIC regularization to disentangle features. During training, CFIPipeline remains frozen, while feature extractors and classifiers optimize jointly, promoting robust causal feature learning. This design effectively separates intrinsic forgery cues from domain biases, enhancing cross-domain generalization.
Key Results
- On four public datasets, G2VD-CLIP achieves over 90% overall accuracy, outperforming baselines by approximately 25 percentage points, with F1 and AUC improvements of 0.194 and 0.104 respectively, using only 10% of training data. The method demonstrates superior robustness across diverse sources, especially in challenging cases like MuseV and SVD.
- Across different backbone architectures (CLIP, XCLIP, Mamba), G2VD variants consistently surpass existing methods, validating its adaptability. Ablation studies confirm that counterfactual sample construction (CFIPipeline) contributes most significantly, with causal disentanglement further stabilizing performance.
- The approach shows strong source-level and dataset-level improvements, indicating broad generator coverage and effective domain shift handling, crucial for real-world applications.
Significance
This work addresses the core challenge of domain shift in AI-generated video detection by explicitly modeling and disentangling causal forgery cues from domain biases. It provides a theoretically grounded, practically effective framework that enhances robustness against unseen generators, advancing both academic understanding and industry deployment of reliable deepfake detectors. Its causal perspective offers a new paradigm for content authentication, crucial for safeguarding digital information integrity.
Technical Contribution
The paper introduces a novel combination of counterfactual sample generation via VAE, frequency-domain alignment, and HSIC-based feature disentanglement within a causal modeling framework. This synergy enables the model to learn invariant, causally relevant features, outperforming traditional domain adaptation or augmentation methods. The design allows flexible backbone integration and demonstrates superior cross-domain detection capabilities, setting a new standard in deepfake detection research.
Novelty
This is the first work to embed structural causal modeling into AI-generated video detection, explicitly separating causal forgery cues from domain-specific biases through counterfactual intervention and feature disentanglement. Unlike prior methods relying solely on frequency or feature alignment, it leverages causal inference principles to achieve robust generalization, representing a significant conceptual leap in deepfake detection.
Limitations
- The approach depends heavily on pre-trained VAE models for counterfactual sample generation, which incurs high computational costs and may limit scalability. Its effectiveness on highly similar or sophisticated generators remains to be fully validated.
- In extremely complex scenarios with multiple overlapping biases, the independence assumption enforced by HSIC may not fully disentangle all confounding factors, potentially affecting detection accuracy.
- Real-time deployment might face latency issues due to the additional inference steps involved in counterfactual and causal feature processing.
Future Work
Future research could explore more efficient counterfactual generation techniques, such as adaptive or online methods, to reduce computational overhead. Extending causal disentanglement to multi-modal data (audio, text) could further improve robustness. Additionally, integrating meta-learning or self-supervised strategies might enhance adaptability to new, unseen generators, pushing towards truly universal deepfake detectors.
AI Executive Summary
The rapid advancement of AI video synthesis technologies, exemplified by models like Stable Video Diffusion and CogVideoX, has revolutionized content creation but also introduced significant security challenges. Fake videos, often indistinguishable from real ones, threaten information integrity, privacy, and societal trust. Traditional detection methods, relying heavily on domain-specific cues such as generator fingerprints or compression artifacts, struggle to generalize across unseen sources, limiting their practical utility.
In response, this paper introduces G2VD, a novel framework that leverages causal inference principles to enhance cross-domain robustness. Central to G2VD is the counterfactual intervention pipeline (CFIPipeline), which constructs synthetic samples by reconstructing videos through Variational Autoencoders (VAEs) and aligning their frequency and pixel domains. This process weakens the spurious correlations between domain biases and authenticity labels, enabling the model to focus on intrinsic forensic cues.
Building upon this, the authors formulate a structural causal model (SCM) that distinguishes between causal factors—representing true forgery signals—and non-causal, domain-specific biases. They design a dual-branch classifier, with one branch capturing causal features and the other modeling domain biases, enforced to be statistically independent via HSIC regularization. During training, the model jointly optimizes classification loss, domain bias separation, and feature independence, resulting in a disentangled representation that generalizes well to unseen generators.
Extensive experiments across four public datasets demonstrate the effectiveness of G2VD. It surpasses state-of-the-art methods, achieving over 90% accuracy in challenging cross-domain scenarios, with significant improvements in F1 and AUC metrics. Ablation studies confirm the critical roles of counterfactual sample construction and causal disentanglement, validating the theoretical foundation.
Overall, G2VD offers a robust, theoretically grounded solution to the pressing problem of AI-generated video detection. Its causal perspective not only enhances performance but also opens new avenues for research in content authenticity verification, promising safer digital environments in the era of hyper-realistic AI media.
Deep Dive
Abstract
Rapid advances in AI video generation pose increasing security risks and call for reliable detectors with strong cross-domain generalization. Although existing methods perform well under in-domain evaluation, their performance degrades substantially on unseen generators. A key reason is shortcut learning, where detectors rely on domain-specific bias rather than intrinsic forensic cues. To address this issue, we propose G2VD, a generalizable AI-generated video detection framework based on counterfactual intervention and causal disentanglement. First, G2VD introduces a counterfactual intervention pipeline (CFIPipeline) that constructs counterfactual samples through VAE-based reconstruction and subsequent frequency-domain and pixel-domain alignment, thereby weakening spurious correlations between domain-specific bias and authenticity labels. Building on this intervention, we further design a causal disentanglement classifier that combines two domain-anchored branches with complementary objectives and a constraint based on the Hilbert-Schmidt Independence Criterion (HSIC), encouraging the causal and non-causal representations to capture intrinsic forensic cues and domain-specific bias, respectively. Experiments across four public datasets demonstrate strong cross-domain performance and consistent gains over baseline methods. In the challenging GenVidBench setting, G2VD achieves over 90\% overall ACC, with improvements of 0.194 in F1 and 0.104 in AUC over comparable state-of-the-art methods, while using only 10\% of the available training data. Code is available at https://github.com/DMOSCAR-98/G2VD.